Repository navigation
Pool composition review: 12h audit of pool membership, best-per-pool choice, and right-sizing #1356
Description
Purpose
Drew (2026-10-27): a regular check on quotas and route pools — whether the latest models are incorporated, whether the best model per pool is chosen, and whether pool members are right-sized against each other and against the pool's speed/cost/performance requirements.
The fleet already has an hourly steward (balances burn across pools as composed) and daily pool-intelligence (AA market data + our measured speeds + add/retire flags). Neither owns membership: twice a day, this review audits whether each pool's composition is still correct and hands the steward changes that are ready to land.
The three checks (every run, every pool)
Pools in scope: scout / workhorse / specialist / summit + *-plan variants in the canonical config (developing-today/code .opencode/opencode.jsonc), plus OpenChamber's routing categories.
- Latest-model check — newly released models (from pool-intelligence's AA data + provider catalogs) missing from the pool where they'd belong, named with AA standing vs current members, plan/quota home, and the member displaced or joined.
- Best-per-pool check — is each pool's current lead the right choice given AA values, the pool's quota window + reset time, AND our historical per-model/per-effort statistics from real sessions (babysitter log, session store, meta-reviewer scorecards)? A lead that wins on AA but loses on our sessions — or draws on a window about to reset while an equal model sits on a fresh window — is a misalignment, with numbers.
- Right-sizing check — members matched to each other and to the pool's job: measured speeds differing ~3× inside one pool doing the same work, a summit-class model on mechanical work, a fast-cheap member carrying work its success rate says it shouldn't, or interchangeable members where one burns a scarcer window — each named with numbers.
Outputs
- A review section appended per run to
hidden_files/pool-composition-log.md(per-pool verdicts; aligned is a verdict too). - Ledger-ready change proposals handed to the steward (pool, change, expected measurable effect, evidence). The review NEVER edits config itself — changes land through the steward's PR path.
- Add/retire flags filed as their own issues when they clear the quantified bar (same pattern as pool-intelligence).
Acceptance criteria
- The
pool-composition-reviewjob is live at a 12h cadence with its instruction file in the goal workspace. - The first review covers every pool in the canonical config plus the OpenChamber routing categories, with per-pool verdicts in the composition log.
- The first run produces at least one steward-ready recommendation, or a verified-aligned result per pool with the numbers that show it.
Filed by Muse (operator) 2026-10-27. Sibling functions: steward (#1339), meta-reviewer (#1340), pool intelligence (#1341), controlled tests (#1342).
Activity
developing-today commentedon Oct 27, 2026
CODE-456
dezren39 commentedon Nov 2, 2026
🤖 Agent note · state: picked up
7 remaining items
dezren39 commentedon Nov 3, 2026
agent-note:v1
state: triaged (coverage pass)
Coverage pass 2026-11-03: reviewed in the fleet-wide per-issue pass (drift / overlap / already-handled / stale / blockers / ordering). Verdict: still live; queued in repo-health-code-pilot as triage at priority 40 (priority/normal). If this is already handled at current main, the triage closes it with evidence; if it needs splitting, the triage files the split.
Found via: coverage pass 2026-11-03
dezren39 commentedon Nov 3, 2026
🤖 Agent note · state: picked up
dezren39 commentedon Nov 3, 2026
🤖 Agent note · state: done (triage — closed as established)
Triage verdict — #1356: ESTABLISHED, acceptance met, close on evidence
Re-checked fresh 2026-11-03 (issue OPEN;
gh pr list --search 1356→ no PR/superseding work; #4606 CLOSED 2026-11-02T22:30:32Z; PR #1520 MERGED 2026-10-28T00:00:56Z at479c575; #1394 OPEN; #2306 OPEN). Prior triage (task 1337, 2026-11-02) found establishment complete and kept this open only pending the cadence check. That check (#4606) has since closed on direct scheduler evidence, so the last condition is satisfied.Acceptance criteria
pool-composition-reviewenabled,interval@12h, 0 consecutive failures; runs 6–13 from 2026-10-29 17:55 CST to 2026-11-02 04:55 CST, exactly on cadence, next run scheduled 2026-11-02 16:55 CST. Runs 1–5 (2026-10-27→10-29) are quoted with timestamps in fleet(pool): background tier has zero consumers — name the adoption/repointing step and owner (composition R3, carried 4 runs) #2306.opencode-route/backgroundoutside its own definition). GitHub-visible output stays silent after run 5 because aligned verdicts file no issues — the silence was investigated and explained in ops(pool): verify pool-composition-review 12h cadence since run 5 (2026-10-29) — re-arm if stalled #4606, not assumed.Residual, tracked elsewhere (not blockers for this issue): #1394 is queued/blocked on controlled testing (#1342); #2306 carries R3 until the background tier gets a consumer. Nothing here requires further establishment work, so this tracker closes; the job needs no re-arm (per #4606).